mvrahden / recurrent-js Goto Github PK

Is your feature request related to a problem? Please describe.
The current examples provide a very limited feature set to learn.

Describe the solution you'd like
Please provide examples with more complex datasets.

Describe alternatives you've considered

Additional context
Add any other context or screenshots about the feature request here.

This line:
import { Assertable } from './utils/Assertable';

Introduces this error on certain configurations of webpack, specifically when importing the library in Stackblitz as I am:

Can't find module:
recurrent-js/dist/utils/Assertable (@1.5.6)

Check your import statements & ensure you're importing the correct module names.

Is your feature request related to a problem? Please describe.
Currently the leakyReLU is missing as a common activation function.

Describe the solution you'd like
Add it to:

• MatOps & Graph
• include the derivative function for backpropagation.
  Wikipedia-URL: Leaky ReLUs

Further related activation functions:

• noisy ReLU
• ELUs

Describe alternatives you've considered
none

Additional context
none

Is your feature request related to a problem? Please describe.
There's no option to choose the activation functions per layer.

Describe the solution you'd like
Please give the ability of configuring a custom order of activation functions per layer.

Describe alternatives you've considered
Providing the key activations in the configuration object would allow the choice of activations per layer.
One could even come up with a solution for custom activations per unit.

Additional context

import {Net,Graph} from 'recurrent-js'

..../minifier-js/node_modules/uglify-js/lib/parse.js:196:18:
Unexpected token: name (Graph)
at new JS_Parse_Error

Did that need CommonJS or TypeScript translator module?

Is your feature request related to a problem? Please describe.
The current implementation of stateful networks exposes implementation details to the user of those architectures, such as the previous state and the graph.

Describe the solution you'd like
To resolve this, encapsulate the graph and the state of the network during forward pass, e.g.:

net.forward([0, 1]); /* or with custom learning rate: net.forward([0, 1], 0.3) */

The encapsulation of the graph and the state will bring the API more close to what we see from the stateless architectures.

Describe alternatives you've considered
none

Additional context
none

Is your feature request related to a problem? Please describe.
Add convolutions via filter functions (Kernel) to learn from multidimensional data.

Describe the solution you'd like
See convolutional neural networks (CNN).
A convolutional layer network that breaks the multidimensional data down (as a form of pre-processing) into a vector, which can then be fed into a NN-architecture.

Describe alternatives you've considered
none

Additional context
none

Is your feature request related to a problem? Please describe.
The memorization of previousInput and previousOutput in the FNNModel is currently used to calculate the squaredLoss directly while backpropagating the target output. This is a weak and error prone implementation.

Describe the solution you'd like
Split this step into two independent steps, e.g.:

let output = net.forward(input);
net.backward(targets);
let squaredLoss = net.getSquaredLossFor(input, output);

Describe alternatives you've considered
none

Additional context
none

Where's the library for recurrent-js-gpu? (from NPM)

I'm very curious to see the GPU bindings, as this library appears to have no real deps, nor does the README match up with the 1.4.9 on NPM.

Sidenote: Any docs on implementing/training/testing an LSTM with the GPU-enabled library?

Cheers